Want to join the conversation? Connect with us on Tribe.
Find “Founders in Arms” on Apple Podcasts, Spotify, and YouTube.
Transcript of our conversation with Parag Agrawal:
[00:00.3]
The market today is like people want to buy tools for agents because an agent is defined as a model with a set of tools and people give it a web search tool. I think what I’m excited about for next year is I think more and more of the market will move towards sub agents so people will call us as a search agent to go do with more autonomy, more work than a single tool call, which is very constraining.
[00:22.0]
And the next thing I’m excited about, which is another product we shipped last week is this thing of the web transitioning from pull as it is today to push. So you should be able to say when X happens in the world, call me so that me or my agent can take an action. Hi everyone, welcome to the Founders in Arms podcast with me, Immad Akhund, co founder and CEO of Mercury.
[00:44.1]
And I’m Raj Suri, co founder of Lima and Tribe. And today we have with us Parag Agrawal, co founder, CEO of Parallel. Welcome Parag, thanks for having me. Parag was previously the CEO of Twitter, just before they sold in their final kind of going private acquisition.
[01:02.8]
So lots of fame. But you also just announced a big funding round for Parallel, right? We did a week ago. Tell us about Parallel. We started Parallel two years ago. The notion that I got obsessed with was that everything we’ve built on the web over several decades is going to become irrelevant when the primary consumer of the web goes from being a human to a, agent.
[01:24.6]
And so two years ago I got obsessed with this notion of like when we building consumer products, how are we building it? We’re thinking about a user looking at a device, what experience to create for them, what inputs they would provide, what to show them and if there’s going to be an agent between us and a human, how you architect every piece of technology on the web changes completely.
[01:47.0]
And so we got obsessed with that idea. And the second thing that really pushed me and the team over the line was like once you get obsessed with it, the very next obvious leap of faith is that agents are going to use the web several orders of magnitude more than humans ever have.
[02:04.4]
Because we’ve been limited by our time and attention, agents are limited by the number of GPUs we can build and deploy for productive purposes or non productive purposes. And that was really fun to now imagine that we’re building infrastructure which is going to operate at orders of magnitude bigger than web scale as we think of it today.
[02:27.9]
Can you give us an example of someone using Parallel Because I guess that sounds like a great vision statement. I’m, trying to put it in a practical use case. So our products are APIs. And if you think about anyone building an agent that needs to access the web.
[02:43.3]
So think of the most popular agent that most people are aware of and familiar with are coding agents. So you have your coding agent occasionally. Most of the time your coding agent is looking at your code, your repo, editing files, running commands. Occasionally it needs to go out to the web to find some information, it needs to read some documents, it needs to get out, it’s stuck.
[03:03.2]
Or it needs to do competitive analysis because you’re doing product work. In those times, these coding agents use our APIs, in many cases to search and fetch content from the web. That’s a very simple an agent we’re familiar with just using our APIs.
[03:20.2]
You see this in anyone building a sales agent, anyone building an agent for finance, someone doing insurance underwriting. So to go back to the coding agent, let’s say you include some library that does something, and the agent will then like search the web to say, for this library, give me the docs, and then you’ll give the URL.
[03:41.0]
Do you also then like go scrape the URL and turn into structured data for the agent to absorb? Exactly. So we gave the difference between a human and agent and you’ve latched onto it, is that you want to give the agent the exact right content in its context so that it can take the next step successfully.
[04:02.3]
That’s the entire goal here. So you don’t make the agent take many steps. If you get the agent to tell you a more precise query than keyword searches, and you produce output tokens which are sufficient and dense in order for the agent to be able to do what it wants to do next.
[04:18.8]
And that’s essentially what we optimize for. That’s pretty cool. And yeah, I guess when you started, agents were pretty new things and everyone was probably still in the form of thinking about ChatGPT and these kind of chat interfaces.
[04:37.1]
Did you think this is the way it was going to go? Or like, you know, you kind of build this generic tool and you’re kind of just seeing these use cases develop? No, I think we made a bet that this is the way it was going to go. I think you have to make some opinionated bets that to you appear very likely when you go about building something that takes a while to build.
[04:57.5]
So you kind of have to build ahead of the market sometimes. And in this case we had the conviction that it would like I was one of these, between Twitter and this company, I was just sitting there hacking on open source. There was this Baby AGI thing and AutoGPT. There are a bunch of these open source agent builders and the models were not good enough but people were hacking together agents at the time and you could feel them.
[05:20.8]
And if you believe that models get better you could feel that these things will become useful. It was vibes, it was like empirical. But also it was clear that that had to happen if AIs were going to be any good.
[05:37.8]
Do you think there’s a world where there’ll be like thousands and thousands of agents? Like there’s thousands and thousands of apps or I guess hundreds of thousands? Or is it that like there’s a few SDI agents, a few coding agents, like there’s like a few big use cases and there’s not that large a long tail. So let’s separate the definition for a second.
[05:55.5]
Right, agent. How many agent companies will be there or how many people will sell agentic products? A separate question from how many agents there will be. I do think there will be a lot of specialization agents and a lot of work will get done by multi agent systems. So I think we’re going from chatgpt to an agent with tools, to an agent with subagents, to more interesting organization primitives of agents together.
[06:22.0]
But if you’re getting at market of agents and what that will look like. I have my takes but I’m actually curious to hear how you’ve seen it. You’re users of agents, I’m guessing it’s hard to play it out fully.
[06:38.6]
The question I always have and I guess it goes back to how many companies are there versus how many. I think there’s a world where companies like Mercury end up building a ton of agents. But we just use OpenAI and anthropic APIs because they just get advanced enough and you don’t have to specialize that much.
[07:00.2]
But the number of things that will be agentic will be very high. I just don’t know if the number of companies that do it will be in the hundreds of thousands or whether there’ll be a few companies that have horizontal tools for it. My take is on the agents which are broad based for consumers to use.
[07:20.8]
I think there can only be a few. Right? It can be like single digits is my opinion. Like consumer broad based generalist agent. I think there will be niche agents which are very specialized for people who only like very 1% of the population like cares about something and is obsessed about something and they will have their niche agents.
[07:42.9]
I think the enterprise side is much more wide open in my mind. I think it could shake out in many different ways. I think on one extreme it could shake out a little bit like the way the cloud and infrastructure story has, which is like a few on the other world. It could shake out the SaaS story had which is like every company is like this bespoke weird thing and ends up having its own stack and someone sort of built a better ish solution for them.
[08:07.9]
You can’t be confident which way it will trend. What are we seeing right now on agents? Like what what is the current like most popular use case that you guys are seeing? For agents today? Yeah, no.
[08:23.4]
So for I think broadly agents are successful in many areas. We only care about agents where there is real value on the open web that these agents need to unlock. Right. So the kinds of agents where we see a lot of value is like any application area of deep research, if you may.
[08:41.9]
So if you’re users of like OpenAI or Cloud Deep research products, those things are really starting to work. In many cases, consumer is the other product we’re familiar with. But like in the enterprise like for finance, for insurance underwriting, for like anytime you need to go like competitive intelligence and And A due diligence and I’m going to build a PE firm, we’ll go do a bunch of research and analysis and pull data, even scientific research.
[09:11.0]
We’re seeing a bunch of now companies show up trying to build agents for doing scientific research and pulling insights from different disparate papers and corners of the web together. So on the high end of people really throwing large amounts of computer, we are seeing that, we’re seeing a lot of scale also in these let’s call them data enrichment use cases.
[09:37.1]
So people will go and take their entire book of business. Let’s say you’re a bank and you have given out credit to a lot of people in let’s say small businesses and you have some ways of knowing signals of credit risk that manifest based on public data.
[09:57.6]
So now people will go build agents to monitor them and make them actionable. People will, when their sales involved people do all kinds of creative things. We see a lot of creativity in people finding everything about their customers or prospective customers and target them in interesting ways or prioritize in different ways or assign accounts to different A’s.
[10:17.5]
So same thing in Recruiting. Right. Like I think you want to find, you know, you allow people to use creativity. Like if I have a hack, which is like, I’m sure Immad you have a hack on like what kinds of people and what kind of backgrounds you think have disproportionate merit.
[10:34.8]
When you’re able to hire them, people will go build an agent to essentially go build the list of people that they should be calling. Do you think we need a new business model for like content on the web? Because I guess a lot of content right now is free because humans look at ads and you know, maybe they’ll click on it and buy something or do something.
[10:54.9]
Whereas the way you describe it, it’s going to be agents kind of browsing the Internet, and like str, you know, never looking at an ad. So yeah, what’s the, Is there a new business model or it’s just like information wants to be free and like there’s nothing to do can I do about it? I think we’ll need new business models.
[11:11.8]
I really think people need new business models. I think the default trend line, right, is that it gets framed as a zero sum game. So people are like AIs are stealing my data to create value and keeping all of it and I get nothing.
[11:27.5]
Right. And the natural reaction to that kind of a thing is that you just try to block. Right. And it’s kind of a value destructive but also full serend of sorts because you don’t always succeed in blocking and you in fact succeed in blocking all the good actors and not the bad actors.
[11:47.2]
And so it’s this weird, dynamic that plays out. So I do think a new business model needs to get created to incentivize people with good content. I mean that would be like a cool thing for you guys to build. Is that something you’re thinking about like the business model? The plan is to figure out the future of the market.
[12:03.0]
Yeah, because like Reddit has these deals with, you know, the big, big companies. But like for any, anyone that builds an agent, there’s no way they can like go to Reddit and like a custom deal for every single agent. Exactly. I think that is where the future is and we have some ideas, but I don’t claim that we have the answers because this is like a, inventing a new business model isn’t like a casual thing one can do.
[12:28.3]
But I do think there is a need for it and I think there is a possibility for it. The reason for the possibility is I think AIs are going to create over the web sufficient value that the pie is going to get meaningfully larger, that there will be a way to split it up that makes sense for all parties.
[12:47.7]
So if you don’t have a major value creation hypothesis, it’s kind of hard to create new business models when you have a real value creation hypothesis. By bringing people together, bringing participants together, like a publisher and an AI company and a customer, you can actually reasonably slice up the pie to make it worth everyone’s while.
[13:05.2]
Now the question is, can we do this at web scale, programmatically? Open market, that’s the hard part. On any individual deal basis, you can play out this hypothesis and figure it out. To do it scalably is the crazy hard problem to go solve.
[13:21.1]
Yeah, it seems like maybe stablecoins could play a role there. Maybe. I don’t think you need it. So I mean the benefit of the stablecoin is that you don’t need a trusted party. Right? Like you could have an agent pay, pay a publisher without like, someone in the middle to like take the money from one place and send it to the other place.
[13:45.3]
Yeah, no, I think that is the benefit. The question of do people trust a trusted party more or do people trust a stablecoin piece of code more? And I think there is. It goes both ways. So I think you do need a center of trust of some sorts of through some set of principles.
[14:03.3]
It is unclear. Yeah, I think it depends how much extra stuff you need on top of it. Like if you need dispute resolution, if you need like fraud, like all of those things end up like eventually needing like a payment company in between. What were you going to say, Raj?
[14:19.4]
Yeah, I mean, I was trying to think. I mean there’s a lot of precedents for this type of thing to some degree. Right. So I mean, I think we’re basically talking about like a picture payment structure or a business model around tokens. Right. Like, so agents would pay per token or pay, you know, some fee per type of token.
[14:36.2]
Right. For. To a publisher or a creator, for, for their content. And, and maybe a certain class of content and you know, that are gated by paywalls and you know, I think, you know, crawling the web used to be free, but we’re talking about now making it not free.
[14:56.1]
Right. And like there’s going to be you know, I think there’s going to be some types of content that are going to be behind gates, otherwise they won’t be accessible. I Think a lot of people will just turn them off because their data will have more value inside the moat versus outside. Right. I think crawling the web is all. Crawling the open parts of the web is free.
[15:13.9]
Crawling the gated off parts of the web is not free. That’s where things are today. Now the question is, will more and more stuff end up in the gated parts of the web or, remain in the open parts of the web? Right. And we would like more and more stuff to remain in the open parts of the web.
[15:29.7]
And the difference between, like. I think there is a difference between content that is paid for but accessible for everyone versus content that is only accessible to a few because we have a business arrangement.
[15:46.0]
Right. Which is. So I think that is a subtle difference, which is a difference between open and closed in my mind. Right. Payment is a different dimension. And is it accessible for everyone that has the same market, mechanic to pay? There’s a different dimension. Yeah.
[16:02.7]
I think the analog, I think about Spotify, right? Like Spotify, there used to be this thing where you used to buy songs and you had to buy it. You know, each consumer had to go buy from the store. Now it’s like, every time you listen to a song on Spotify, they get paid. Right? So I think, Is that the analogy I’m sure you’ve been thinking about a lot?
[16:21.6]
Yeah. No, actually, that analogy has many layers to it, right? Like there was like piracy and like Napster and like all of those things. And it felt like, okay, why would anyone pay if you could get it for free? There was a, like, layer of like, DMCA that was helpful in the journey.
[16:41.8]
There was, Spotify that figured out the new business model. Now everyone wasn’t happy in this set of movements and there were winners and losers and all kinds of things happened, but ultimately you took this wild, wild west of pirated content and ended up in a, at least in the west now pretty much, most content being used is licensed when it comes to music.
[17:09.8]
And I think it was a journey. I actually don’t know the Spotify story as to, like, how well designed orchestrated it was and planned it was, versus seeing problems, seeing incentives and going and solving them step by step.
[17:24.9]
But that’s a great, great analog if you think about it. I heard a good reason for why Shopify succeeded. It was that it started in Europe where the labels didn’t care that much. So the Spotify, not Shopify. Yes, Spotify. God damn it.
[17:40.7]
There’s this funny, like, tweet from to Fixing up anyway? Yeah, that it started in Europe where the labels just didn’t make that much money. So they didn’t care that much about the experimentation. And it got to a certain scale in a certain scale in a mass scale space where then they were able to go like, oh, actually this worked.
[18:03.5]
Why don’t we do in the US So maybe there’s analogy here where you could pick a vertical where there’s not that much money, so people willing to experiment or it’s like a tertiary thing, and then get scale. I think that is interesting. We’re just starting to now build that second part of the business.
[18:20.3]
So far our products are all APIs. We sell to AI companies or agent companies or people trying to use AIs to make their businesses better. We have recently hired people and started now partnering with some publishers and figuring out our thing.
[18:37.0]
The core hypothesis we have, which is I think a little bit counterintuitive for people, is that we want to figure out a way of doing differential pricing. Yeah, that’s tricky. I really think the only reason, like businesses might pay more than consumers. Exactly.
[18:52.1]
I think if you try to pay the exact same amount for every read of a data point, there is no way to clear the market and make things work. And so you have to figure out how to do some interesting scalable differential pricing.
[19:09.0]
That’s what ads do. Well, right. Like it’s crazy that like, there is such a nice market mechanic with ads to pay for different queries and different users end up making a ad business very different. Like some are 10% of queries are positive, the rest are negative, negative.
[19:28.9]
But it’s still a high margin business. Tell us about this fundraise, Barag. Who did you raise from? How did it go down? Is this your first company, by the way, or did you do a company before Twitter? Dude, Twitter was my first job. Oh really? Out of grad school. Out of grad school I got hired. How long were you at Twitter?
[19:44.7]
11 years. Wow. Out of grad school, I was hired as the most junior engineer at Twitter. Like you and IC1 or whatever. Yeah. I don’t think Twitter was early enough that there weren’t serious leveling plans or leveling ways.
[20:02.7]
The fundraise, this was our second fundraise. Parallel is about two years old. We started last January. It’s interesting, when we were building for the first year, the market wasn’t there. We were selling to, we were saying we have built APIs and our pitch was crazy.
[20:18.4]
We’ve built APIs. They’re really slow, so humans don’t have patience to use them. But agents will, because agents will be asynchronous. And there weren’t, as you can imagine, a lot of buyers lining up to go buy that last year. This year, two things have happened.
[20:34.9]
We have scaled. Now some people were willing to buy that, and we got to work with them and really scale the product offerings. But then we were able to make things fast because our index really got to scale. So this year, I would say we hit what feels like product market fit to me.
[20:53.6]
I don’t know. Like, so when you say your index got to scale, like, you’re caching the search results and like, you’re caching. No, no. We’re crawling the web. Right. So we. The first thing we did when we started was we put a couple of people on, like, going and building a crawler and crawling the web to collect and build a large index.
[21:09.2]
And it takes a while. It’s like a lot of engineering at scale. We wanted to build it ourselves from the ground up because we knew it would be different for AIs, and we had some hypothesis on how. Yeah, okay, sorry. I don’t want to lose this fundraising story. So you feel like you hit product market fit? Like, did the VCs come knocking?
[21:26.5]
Did you go do a fundraise? So, you raised 100 million. Who was the lead? We raised 100 million. It was co led by Kleiner Perkins, Mamoun and Index Shardul. So they were co leads. Shardul was already an investor from the seed. The.
[21:42.6]
Yeah. So the way it went about was that we felt like now the market is arrived. We’d been building like, quietly and in stealth and selling privately. There was not really a real website. We weren’t really marketing the product. We weren’t really doing very much because we didn’t think the market was there.
[21:58.7]
And then as soon as we figured the market was there and the product was there, we changed into, let’s go higher, let’s go make noise, let’s go launch the product, let’s go talk about the product. And we shifted into a completely different year. In August is when we first launched the product publicly.
[22:18.0]
And I think that created both a wave of customers, but also investor interests, and we were shifting into this more aggressive gear. So it made sense for me to go fundraiser. So it was some combination of inbound VC interest and me thinking that this was the point.
[22:36.8]
Nice. Why kind of build quietly like, why not just launch earlier to garner interest? Just as a strategy thing. I mean, how do you even find customers if you’re building quietly? The thing for us was our product was designed to be an API because we want to be a web player and we want to be horizontal, so we can’t go solve problems that require deep, intimate understanding of a specific agent anywhere.
[23:03.1]
Right? So now almost by definition, if you want to be horizontal and on the web, your product is an API. So we took that for granted because that’s what got us excited about this in the first place. Now, if your product is an API, you don’t get to innovate constantly on the surface of the API once you take on a large number of customers.
[23:24.5]
So we, I, got advice from Patrick Collison early that like, if your product is an API, before you give it out and have a lot of people rely on it, better have confidence in the API shape and the surface, because you’ll have to maintain it for a long, long, long time.
[23:42.0]
Right. So when we initially started, we had a rule that we will only have small number of customers who are very, very close partners, wherein we have a slack where we feel comfortable and they feel comfortable that if we make a breaking change to the API and call them that in a week or two, they will go make some changes and deploy them in their code base and work with us and it will be a distraction.
[24:08.0]
They’ll be annoyed. But that’s the kind of bar to be an early partner because we were guessing what the API shape should be and working with customers to figure it out. And if you take on a large number of customers, kind of, take on too much friction for yourself to be able to sort of then change your mind or iterate fast or change the API shape or the level of abstractions.
[24:34.2]
Yeah, makes sense. Parag. This is really cool. So where are you at today in terms of, You sound like you’re getting product market fit. Your customers are mostly these AI companies. What’s the stage of the product now in deployment? You know, any, any sense of scale in terms of how many queries you’re handling or, you know, how many tokens or anything like that.
[24:55.2]
I think we, the only public information we’ve shared so far is when we launched in April, we already were doing millions of queries a day. The interesting thing about our product offering is that we are shipping APIs that operate at various levels of abstractions. So one of the crazy things, and we shipped a crazy expensive API of today you should try it by the way.
[25:15.5]
But our cheapest API is probably like under half a cent per API call and until yesterday our most Expensive API was 2 and a half dollars per API call. And as of yesterday you can make a single agentic API call and like spend $50 on it.
[25:30.9]
Wow, what does it do? Or hundred dollars on it like sorry, what does it do? It generates a full database for you so you can go say as Mercury. Find me every startup that has fundraised in the last three months as is in the Bay Area and has a founder who was previously is a first time founder and raised from one of these five investors that I know really well.
[25:58.6]
You could run this query and you’ll get like 200 answers. You give it like a natural language problem like that. Yeah. Wow. You give it a natural language query. People do these kinds of queries with deep research historically or products like that and I think you get like a, you get like eight obvious answers that you already knew.
[26:15.0]
You use us and you’re going to get like 180 answers and you can build a full database from it. Wow, that’s powerful. Yeah, it’s just like an army of agents going and doing work for you. Is it like is actually fixed dollars 50 or is it like variable? Depending on the query might cost $200.
[26:33.4]
It’s variable. So you get to say like I want to pay this much per output that satisfies my constraints and parag. So this is an interesting use case you just mentioned to Mercury. Right. Like okay, building this database of like and you said sales before is like something you’re seeing a lot of growth in.
[26:51.5]
Are there any kind of other major areas you’re seeing like hey, these are the applications that people are using this stuff for? Yeah, no, I think in finance and people in the investing business whether it be like hedge funds or PE funds. Like you see a bunch of we see some action in recruiting, we see action in coding agents obviously just because they’re so pervasive.
[27:15.5]
We see use cases in, let’s call them productivity software. General if you think of a piece of software you use at an enterprise internally, and has some system of record and you need to connect their agents to the web, they can use us to do that.
[27:31.9]
We see I don’t know any category of knowledge word. Like people are doing like menu extractions from the web. People are building their equivalent of zoom info which is proprietary to their use case. From the web. People are doing claims processing for insurance or people are building like a data set for doing some financial modeling as a consultant.
[27:58.8]
It’s very, very, a lot of government data is being gotten to do RFPs. I’m going to do payroll in a compliant way. I need to know for every county or every state some set of data which changes every few months.
[28:14.2]
I need to keep that update in order to do compliant payroll. It’s very, very broad based. And you’re serving the people who are making the queries, not who are receiving the queries. Right, we are serving people building automations around these.
[28:29.9]
So we are just an API. So I’ll give you some examples. Right. I think one of our customers has built their own insurance underwriting workflow. Right. So they will then call our APIs.
[28:48.4]
They’ve architected a workflow system that uses our APIs. Now another one of our customers and again there’s different layers of customers. Like we’re integrated into a bunch of low code, no code tools that people might use. Like if you think of N8N or Lindy or Gumloop, there’s a bunch of such tools being used all over where business people will come build their, call them agentic workflows through various modalities.
[29:16.8]
And when that workflow needs to research the web or pull data from the web, they will use our APIs because there’s like a node built using our APIs in those solutions. So we are the very infra heavy part of it and our product is an API and people make this capability accessible to all kinds of users through different ways of integrating them into agents or solutions.
[29:41.6]
That’s cool. And what’s the plan? Like what do you think about next few years? Like do you see yourselves as like architecting this business model or are you going to be like working with other players? And what else do you have in, in store in terms of evolving the company? Oh, so I think the agentic web has barely started.
[30:01.1]
Right. Like there are like still the primary business of the web is humans. These are all, everything I described so far is an edge case. Like almost no one is. It’s not happening yet at scale.
[30:16.9]
So if you think about we are very early in our journey. Like our systems, in some sense we have the most advanced systems and in some sense they are really terrifyingly bad. And so I think it’s very early. The thing I’m most excited about next is the market today is people want to buy tools for agents because an Agent is defined as a model with a set of tools and people give it a web search tool.
[30:42.7]
I think what I’m excited about for next year is I think more and more of the market will move towards sub agents so people will call us as a search agent to go do with more autonomy, more work than a single tool call, which is very constraining.
[31:00.5]
And the next thing I’m excited about, which is another product we shipped last week, is this thing of the web transitioning from pull as it is today to push. So you should be able to say when X happens in the world, call me so that me or my agent can take an action. So we shipped a really early alpha product about this and that’s my favorite direction.
[31:24.0]
Now we will see when it becomes good enough to be useful in very real enterprise use cases. Right. So, so you could tell him that like anytime like someone updates their profile to like founder or something and he’s like calling them, it’s like you should set up a Mercury, account.
[31:40.9]
Exactly. So I think like when something changes in the world as visible on the open web, that is actionable for me, call me. That’s the whole point of that product. That’s pretty cool. What’s been like the most surprising part of starting a company? Parag. I mean you, you were at a very high level position at a pretty, pretty big company just before this.
[32:00.7]
So you’ve seen kind of, you know, a pretty broad scale. But yeah, as an entrepreneur, what was like something you didn’t expect? I got good advice. I think, I think like almost everything that works at scale, at post PMF is the wrong thing to learn a lesson from without scale, before pmf.
[32:22.5]
So almost everything is probably, that you’ve learned is probably the wrong answer. And Twitter had this like super crazy strong product market fit. Right. Like it just grew like an insane amount without like trying that out. Yeah. So I think Twitter almost was like crazy product market fit a scale problem.
[32:40.1]
Any idea you could test empirically. Like, you didn’t have to be extremely opinionated about the future because empirical data, user feedback was so readily available. Like you didn’t have to work to get feedback, you didn’t have to be creative to get feedback.
[32:58.1]
Right. You just had to observe problems and solve for them. In the zero to one journey, you have to be very like the first thing that we had to do was we had to disqualify almost all use cases so that we were set up to build something that we think would matter in the future. Right.
[33:13.5]
So for the first year, we almost had to not be responsive to 90% of things customer asked us on hearing our pitch, and we would say, we cannot do that or we won’t do that. And we had to zoom in on, the places where we thought that we could build something special and ignore where the market or the demand was in the moment.
[33:36.1]
And that was really counterintuitive to me. But where I cannot articulate why we have the right to win. I don’t want to go build, even as a startup. And you also want to build the things that will matter in three years time.
[33:52.3]
They’re the horizontal things that you were talking about that are reusable and. Exactly. No, we are in the classic mode of like, this market is tiny right now. We want to be early and great in this market, and we’re just betting, putting all of our eggs in this basket, that agents on the web will be a big, big, big deal, and this market will be thousand x, a million X of what it is today in a few years, and that we will have some substantial share of this market.
[34:23.4]
Yeah, makes sense. Any other surprising learnings? No, I think I’d never fundraise before, which is kind of interesting thing to learn. I got to speak to a lot of founders and heard a lot of opinions along the way, and everyone does it differently. Everyone has a different philosophy. Forming my own.
[34:41.7]
Seems like you figured that bit out pretty quick. I don’t know yet. I don’t know. We’ll find out in the next one. The thing about fundraising is you only have to fail once for it to be a real problem. Well, in my last company, I failed to fundraise, like, almost every year for like four years.
[34:58.4]
So the alternative is you eventually become profitable, and that’s. That’s the other way around. Yeah. I think I would rather figure out how to build a great product and then never be great at fundraising. If I could have my way. Yeah, that would be great.
[35:14.7]
That would be the utopian. I don’t know about you, but not anymore. But previously, at least, like, fundraising was like, the worst part of running a startup. Like, I like. I like talking to customers and selling. Even if I had no, like, building a product. Like. Yeah, I guess firing people is also probably like the good contender for the worst part of running a startup.
[35:35.2]
But between the two of them, like, I’ve, Yeah, even now, I don’t love fundraising. Wait for you. Fundraising and firing people are at the same level? No, I think fundraising is Much worse than firing people. I would say was like, at Mercury, we’ve had an easier time and I probably got better at it.
[35:55.5]
But yeah, at least at a similar level even now. Yeah, I probably agree. I probably agree. Okay. I don’t hate it as much as you guys do. Well, you’ve had such an easy time. For us, it’s the times that you fail when you like, you know, you go out there and you get like 30 no’s and it’s like, you know, you have like your personality attached to the startup success.
[36:18.1]
So like, when you get a no, it feels like someone’s rejecting you and like, you put your heart into it. It’s like really hard. But. But okay, let me tell you my perspective. Like, for me, getting a no during the fundraise just doesn’t hit me personally. Because my mental model of fundraise is I only need one yes.
[36:34.5]
It’s like the, it’s like an amazing game where there is some correlation, but there is some lack of correlation among all the people you speak with. And you only need one yes. On the customer side, though, I feel like ultimately to win, you got to get like all of them to say yes. So like, every customer rejection is, like a dagger for me, you know, But a fundraiser rejection.
[36:58.3]
I’m like, I need one out of n on customers. I need like 80% of them of the eventual market to say yes to me. And so that’s like way more personal. The thing with customers, at least for Mercury, because there’s like hundreds of thousands of them or even millions eventually, at least at the start I was like, if 1% love it, that’s amazing.
[37:19.8]
But I see your point. And it is true. You only need one. I think the thing that I always get caught in, in fundraising is you talk to someone and you’re like, wow, this person really gets it. And I’d love to work with this person. And they’re like, they have a shiny brand and all the stories about, or they take you on a helicopter ride, whatever it is, and you get like, very excited about working with that person.
[37:42.7]
And then you get a no. And it’s just like, I, do feel like you get some emotional investment into like a specific firm or specific person. Yeah, you spend a lot of time and spending a lot of time with some of these firms. And and the diligence process can be, you know, kind of excruciating in many cases.
[37:59.8]
And VCs are trained to like, be very excitable. Like they, they Want to seem like they’re like so into your idea until like the last moment when they But I think perhaps I already have. At least my mental model here is helpful. Almost by definition, a VC wants it is rational for them to be excited about you and indicate to you that there is an extremely high likelihood of them investing at an extremely favorable set of terms for you so that they have the option.
[38:37.2]
Like if they don’t do that, you’re not going to allocate time to them. Yeah. I mean it’s 100% rational from their perspective. Yeah. And so now if you assume that then there is no signal in there. And like as much as, and they’re not being misleading, it is just like it is rational for them to have you allocate time to them, which will happen if they are curious, if they’re excitable, if they’re excited so that they get more information from you.
[39:05.9]
So that they have all of the information. Yeah, I mean that’s all rational, Parag. But we’re kind of humans and we’re like on the other side, you as an entrepreneur, selling your heart out. Right. And you also want to. The game you’re playing as an entrepreneur seemed that you’re very excited about that firm and that you really believe in your idea and you know, you can’t imagine why someone else wouldn’t.
[39:30.4]
Yeah, I agree. Both people are playing like a, like a game and projecting something, but it’s hard to like completely remove the human aspect of it. Yeah, no, I’ll bring it back to customers. I feel that way with customers. Like I think sensing that you’re very likely to have a customer work with you and then that not working out is the one that really hits me more than other things.
[39:55.3]
Like I know because I can imagine that being hardest. Yeah. Because there, there isn’t that incentive like, like customers have almost truth telling incentive because it’s really more of a partnership when you’re selling an API. And there, if you read the wrong cues, like I’m not discounting what a customer tells me.
[40:11.1]
Right. And there, if it doesn’t work out then is it like, oh, did we, did I not understand their product use case? Did I screw it up? Is my product bad? Like that sends me into a more negative cycle. Yeah, I could see that. Raj, did you also have that with customers?
[40:29.9]
Were you always kind of emotionally attached to winning each one? No, no. The thing is with customers, yeah, they have a truth telling attempt but like you can iterate typically. And I’VE had meetings with customers where like, yeah, they said no, but then they gave me some new ideas for how to improve stuff. And like, And those meetings are always valuable, right?
[40:46.9]
Whereas with the VCs, I. I didn’t feel like the meetings were valuable. I felt like they were a waste of time because, like, you know, because they were called, I think there, there was a word for it, you know, I have to use this cuss word here. Grin, fucking or something. Like, they’re just like. So, you know, that is the right term though.
[41:05.8]
Yeah, they’re like, oh, that’s so great. That’s amazing. Blah, blah, blah. And they’re like. And they just waste your time. That’s the thing. And. And like, Parag, you definitely have not gone through this because you already have a brand, like, attached to you and like, you’re kind of already well known. So, like. But like, you know, Iman and I have to go through this many times.
[41:23.3]
I mean, Iman, maybe you’ve failed for four years. I probably failed for 10. It was very hard to raise for like, presto. So, yeah, it was a bane of my existence. But I mean, I got good at it as a result. So I’m like, yeah, you know, yeah, I think I’ve been lucky so far. I will come around to your perspective perhaps in the next couple of fundraisers.
[41:42.8]
Hopefully it’ll be up into the right. You’ll never have to deal with it. There’s no such thing. No such. What’s your. What’s your view on, like, you know, are we in an AI bubble? I mean, you, you see it at the ground level. Like, you know, are we just. I actually have a hard time deciding sometimes whether we’re like just getting started or whether we’re like in a late stage bubble.
[42:00.7]
It’s like a weird moment. So, okay, perhaps at the least controversial opinion here, I think there is. It’s not a bubble. When we are like eight years from today, this moment will not feel like a bubble.
[42:19.1]
Is it possible between now and eight years from today? At some point there is like a big downswing because people did stupid things. You know, in the dot com bubble, it took 14 years to recover. I think here it’s like more five to eight, in my opinion. So I do think time is moving faster.
[42:35.8]
Like everything happens faster. So you don’t need that long in my intuition. So I think instead of thinking of this as the bubble, you have to think of that downturn as the overreaction in my mind. But you do have to think of like, okay, there is exuberance, there is people willing to experiment.
[42:56.8]
So you do see the signals of what might feel like a bubble. Right. Like people have, like, oh, I need to use AI Instead of saying I need to solve problem X, people are saying, here is a solution. Where do I use it? When you see behaviors of that kind, which we’re seeing all over the market right now, there are risks because we’re generally very, very bad at saying, I have a solution.
[43:23.0]
Let’s find a problem to apply it to in general as people. And so we are going to do bad things. Now the question is, will the good things continue to deliver enough value for us to have the luxury of making these mistakes constantly and wasting capital and misallocate capital, or will they not?
[43:44.3]
And I think that’s what will decide whether we enter into some negative moment or not. But the reality is there’s enough good to be done that if we all did it over time the right way, we will get way beyond where we are in terms of positive impact and value creation.
[44:01.9]
And so again, I don’t know which way it will go. So I don’t know if there’ll be a downturn moment. I’m hoping there won’t be, but it’s possible. And part of the funding strategy for us is to be able to absorb that is the idea that you’ve raised a bunch of money and you’re not going to spend it very quickly.
[44:22.4]
Not exactly. The idea is that we will spend it against business goals. So we’ll spend it quickly if we see revenue scale like crazy. Right. But if we don’t, we won’t. So I think it is a somewhat disciplined strategy aligning, spend to market traction.
[44:44.5]
Yeah, I was wondering, because you’re also building this crawler, and I assume you have some pretty expensive compute needs as well, which are not tied to revenue all the time. Exactly. And so I think we really manage and govern that. Right. So that was relative to the, to the capital we have and the customers we have.
[45:03.3]
We try to do math to figure out how much to spend there. So we’re taking a somewhat disciplined approach, but not a non aggressive approach, if you may. So if we see revenue growth, we are going to go spend faster, but if we see any slowdown, we’re going to be much more measured.
[45:21.1]
And so should we do the rapid fire? We have a rapid fire. Yeah, I’m terrible at rapid fire, but let’s go. It’ll be fine. Just, I can start off. What’s the biggest mistake you’ve made in your career? Not shipping a piece of software like today and waiting for like six months.
[45:37.8]
Was this at Twitter or current company? Yeah, Twitter, yeah. Do you have the specific piece of software in mind or you don’t want to reveal? No, it was just like recommend ranking in the timeline. So that thing has compounded ever since Twitter was flatlining. That thing is compounded ever since.
[45:53.0]
And the cost of six months of compounding, it’s just like the cost grows every year. Zoe, what do you mean ranking in the timeline? So Twitter timeline used to be reverse chronological. I went to a new team, worked on figuring out how to build the algorithm.
[46:09.0]
For it, and it took us a, we were almost too rigorous in shipping it. That’s interesting. So this was actually like inflection point at Twitter. Like it was like flatlining, like usage was flatlining and all this stuff until the algorithm was shipped. Yep.
[46:24.2]
And you could have shipped it like it was ready to fire, I think six months, nine months before we actually got it out at scale. And those are six months of compounding. What made you stall? What led you to stalling? It wasn’t just me. I think that was like, it felt like a big fundamental change to the platform and people were slightly risk averse and wanted it to dot their I’s, cross their t’s.
[46:45.3]
And it was. You just did 10 more experiments and like understood every angle of this instead of just like shipping. Yeah, yeah, yeah. That tends to happen unless you have like a very, I guess, decisive CEO or something that can just go like it’s ready go kind of thing.
[47:03.1]
And is this, this is the for you page, basically, right? As it’s become? Yeah, yeah, this is the. And you think that if you did that six months earlier, like the growth of Twitter would’ve just Like even though six months would have made a big difference I think, because if it’s since then, at least until I left, Twitter grew by 15% year on year in terms of usage.
[47:23.6]
Right. And so Twitter would be 15% larger on any given day if you shipped it a year earlier. Right. And so that number is now growing each year. And so now you like every year the regret increases.
[47:39.3]
And that’s I think an interesting life lesson to working in areas where there could be compounding where you’d rather ship early than late. It’s only really bad if a competitor ships it before you and then they compound away from you or something.
[47:55.1]
Otherwise it doesn’t matter too Much. If it’s like, one year delay. No, that’s even worse, though, right? That’s even worse. In my opinion. Like, this is the baseline. Like, even this is such a big regret that if a competitor further destroys you as a result of you being slow, then you’re, like, totally written.
[48:11.3]
Which founder inspires you the most these days? I don’t know. I’m a little bit. I, no longer have individuals as big idols these days. But when I took the job at Twitter, I was really inspired by Satya, and I used to call him the founder, or a re.
[48:28.5]
Founder, or whatever you want to call that. I do think that it’s like. Like when I was taking on this Twitter CEO job, right? It was this moment where I had to figure out what this company was going to become with me and how it was going to be different.
[48:44.5]
And the inspiration I got on someone, like, completely changing a company, a business, a culture, a product, came from Satya. And I thought of wanting to be a founder in that mold of a founder in that moment. Yeah. I feel like Satya, has done a lot of crazy things, and with a lot of confidence, which has been really admirable.
[49:05.7]
Did you have some crazy ideas? Obviously, you had a relatively short tenure, but given time, were you going to go fund a open source, AI Nonprofit or something like that? No. We did fund a thing called Blue sky before.
[49:21.0]
That was before I became CEO even. But I think there were crazy things to be done. I think Birdwatch, as we used to call it, or Community Notes for moderation was something I was building and super excited about. We had this thing called Project Saturn, which was crazy, on how to change the content moderation regimen.
[49:38.5]
We had this notion of how we would change the recommendation system. So, yeah, I think we were going to shape Twitter to be very radically different than what it had been. I didn’t get the full opportunity to go do that, but I think that’s what I took inspiration from.
[49:54.7]
So I think what I’ll say is being a founder is like a state of mind more than I would agree. I mean, I think people, early employees like you are Twitter and like Satya, at Microsoft, like, they have. That is built. You know, it’s kind of gets, built into their personality.
[50:09.9]
Right? Like, they’re. This is who they are. You know, they’re. And I think they can act like founders as well. It’s unlike a hired gun. We’re bring someone coming out from the outside who doesn’t have that, okay. What has been the most rewarding part of your founding journey to date stuff.
[50:28.5]
Working for a customer. Like, every time that happens and you see someone do something surprising happens, like once every few weeks, I want to say, where, like, you see one thing and you’re like, wow, I didn’t think someone would do this. And it’s, like, really, really weird and really, really cool. And it tells you a little bit about what the future looks like.
[50:46.0]
Like, at this company, we’re kind of living in the future because our customers are doing strange, futuristic, creative things. Yeah, that is cool. You can see a trend before it happens. Yeah, I’ve had that a few times. Yeah. All right, maybe the final one.
[51:02.4]
Which current trend of fashion do you think is a passing fad? Current trend or fashion? I think video, fundraise announcements. You think so? Well produced video fundraise announcements.
[51:17.9]
I think that’s gotta go. There was like an, there was an alpha to them when the first person that did it. But, yeah, now everyone’s doing it. Yeah, that has to go. It’s interesting how fundraising announcements is changing over time. Right before it was all like, TechCrunch was like, the rite of passage.
[51:35.0]
Now, everyone’s doing direct stuff. What’s your answer to the fad question? That’s a good one. Yeah. Okay. This one’s probably controversial, but I think this, like,996 thing is a fading. You were going to say 960. I was just talking about it yesterday. I just think, like, I don’t know, I’m a pretty good engineer and I cannot work996 as an engineer.
[51:54.5]
Like, you just can’t code that much. Like, maybe, like, obviously sprints are different. Like, yeah. You kind of grind for a bit, but, yeah, I used to be 21 and I used to do 996, but you know what I do? I’d sit there, like, have my lunch, watch. Yeah. Read hacker news, go on Twitter.
[52:09.8]
Like, I wasn’t really working996. I was just in front of my desk996, because I didn’t have a life. So I think, like, at some point, people realize that, like, actually productive time is what matters more than, like, how long you sit in front of your desk. Yeah, I totally agree. I was going to say the exact same thing. So much.
[52:26.9]
It’s going to. It’s going to pass at some point. But we see these, we see these. You know, things come back in waves. Right. So it’ll pass and then It’ll come back in some degree then, you know, so it’ll happen. I think, you know, this, this idea of Y Combinator and some company or some folks are like focused on in office.
[52:43.7]
Like that’s become the main focus. Especially big corporates are focused on in office. I think Immad and I both run remote companies and I see a lot of benefits to that. And so the pendulum keeps swinging back and forth. But running remote companies can be very, very efficient. So I think there’s going to be a lot of that.
[52:59.5]
Are you remote or in Bersenberg? We started with religiously in person until a few months ago when we started scaling a lot. And now we’ve made a few rare exceptions. But engine product, there’s like very few exceptions.
[53:15.1]
Like we were very in person in Palo Alto five days a week. We were, I was thinking pre product market fit. Like being in person I think is like probably the important, but eventually I think the benefit, like once there’s like we have thousand 100 people, like you can’t even fit that many people like on the same floor.
[53:34.5]
Like it’s not like you’re really in person in person anyway. Yeah, no. I do think that’s why now we’re becoming slightly more flexible. But I think the bias is still a little bit in person. We were so crazy, like so opinionated that we would not often dial into meetings.
[53:51.9]
So if you weren’t in the office you’d miss it. We didn’t use project, manager or task manager. We would write it on the whiteboard and we were like a little bit like early on in the first year we tried to build extreme incentives to be in the office.
